Jul 29, 2008 - However, while the health, social, and economic costs .... effects differ between children (level-2) and between schools (level-3) ...... P.D.N. Simpson, J.D.N. Gossett, C.D.N. Connell, D.D.N. Harsha, C.D.N. Champagne, J.D.N..
Economic, Environmental, and Endowment Effects on Childhood Obesity
Minh Wendt Department of Applied Economics University of Minnesota
Selected Paper prepared for presentation at the Agricultural and Applied Economics Association Annual Meeting, Orlando, FL July 27-29, 2008
Copyright 2008 by Minh Wendt. All rights reserved. Readers may make verbatim copies of this document for non-commercial purposes by any means, provided that this copyright notice appears on all such copies.
ABSTRACT This paper examines factors associated with the childhood obesity phenomenon in the U.S. A national longitudinal dataset “Early Childhood Longitudinal Study, Kindergarten-Fifth Grade” (ECLS-K) that has data for 12,719 children from fall 1998 (Kindergarten year) through spring 2004 (Fifth grade) is used. Two econometric models, a mixed-effects ordered Logit and a random-effects Tobit, are used to predict obesity status and the extent to which a child is overweight (the excess level of a child’s Body Mass Index). The results show that the more time parents spend working and certain non-parental care sources such as at-home childcare (as opposed to childcare at centers) are statistically significant in predicting the likelihood of childhood obesity and their level of excess weight. Endowment factors such as child’s birth weight and race, and other demographic factors such as parents’ social economic status, family structure, and family size are strongly correlated with these two outcomes. Environmental factors such as bed time, computer use, and physical education programs at school are negatively correlated with the level of excess weight but not statistically significant in predicting the probability of being overweight in children. The results of this study will be useful for educators, parents, and policy makers.
ii
Table of Contents 1. INTRODUCTION.............................................................................................................................. 1 2. ANALYTICAL FRAMEWORK ....................................................................................................... 4
2.1 Theoretical model ....................................................................................................................5 2.2 Reduced form estimating equation ...........................................................................................6 2.3 Econometric estimating models................................................................................................7 3. DATA AND VARIABLES................................................................................................................. 9
3.1 Dependent variables.................................................................................................................9 3.2 Independent variables ............................................................................................................12 3.2.1 Parental time ...................................................................................................................12 3.2.2 Environment ....................................................................................................................13 3.2.3 Endowment.....................................................................................................................16 3.2.4 Other demographics ........................................................................................................16 4. EMPIRICAL RESULTS.................................................................................................................. 17 5. CONCLUSION................................................................................................................................. 25 REFERENCES ..................................................................................................................................... 27 APPENDIX ........................................................................................................................................... 33
iii
1. Introduction According to the World Health Organization (WHO), there are over one billion overweight adults worldwide, with more than thirty percent of those being obese (2006). Overweight and obese adults are defined as having a body mass index (BMI)1 over 25 and 30, respectively. For children and adolescents, these terms are defined as having a BMI above the 85th and 95th percentiles (of growth) for a child’s age and sex group. Obesity rates have risen three-fold or more since 1980 in both developed and developing areas of the world. Overweight rates in U.S. children and adolescents 2-19 years of age have more than tripled, from 5 percent in the 1970’s to 17.1 percent in 2003-04 (CDC, 2006). Childhood obesity is either directly or indirectly associated with a variety of health problems such as sleep apnea, type II diabetes, asthma, mental health, and most importantly, greater risk of becoming obese adults (HHS, 2004). The escalating rate of obesity highlights the importance of understanding its causes and consequences. However, while the health, social, and economic costs of obesity are apparent, the mechanisms underlying this problem are less obvious. It is widely believed that parents play a central role in their children’s food choices and activities, affecting the nutritional and physical health experienced by those children. To achieve healthy outcomes, parents have to devote sufficient time and income to care for their children. Given the assumption that both nature and nurture factors affect childhood obesity, this paper seeks to understand the influence of parents’ time constraints on childhood obesity, using Becker’s (1965) economic framework of a household production model (HPM). Particularly, it explores the relationship between childhood obesity and the quantity and quality of the time that parents have for
1
BMI is defined as the weight (in kilograms) divided by height (in meters) squared. Equivalently, the weight (in pounds) divided by height (in inches) squared and multiply the answer by 703.
1
their children, controlling for certain environmental and endowment factors. In other words, it investigates the effect of parents’ time input in a household production of a healthy output, i.e. a not overweight child. The data analyzed in this study were from the U.S. Department of Education’s Early Childhood Longitudinal Study, Kindergarten class of 1998-99 (ECLS-K), merged with two other sets of data: monthly housing cost for homeowners with mortgage from the US Census and weekly average wages from the Bureau of Labor Statistics2. The ECLS-K project is an ongoing study that collects data on a cohort of children’s early school experiences beginning with kindergarten and following these children through high school. The ECLS-K is the first national survey on public and private kindergarten programs and children who attend them. The newest round of data available is through 5th grade in the year 2004. This dataset provides a wide range of data needed to understand children’s health, early learning, development, and education experiences. Data is collected for the child, the child’s parents/guardians, teachers, and schools through child direct assessment, home and school interviews. The ECLS-K survey used a multi-stage probability sampling design. A nationally representative sample of approximately 22,000 children enrolled in 1,000 kindergarten programs during the 1998-99 school year were selected for participation in the study. In this base year, the primary sampling units (PSUs) were geographic areas consisting of counties or groups of counties
2
This is done by using home-based census tracts available from the ECLS-K restricted dataset, grouped together at the county-level and/or state- levels in order to merge with available information from the US Census Bureau and the Bureau of Labor Statistics. Most of monthly cost of homeowner with mortgage data is available at state-level, and wage data is available for over 300 largest counties or 100 metropolitan areas. Therefore, state-level or metropolitan-level information is used for counties that do not have data.
2
selected with probability proportional to size (PPS) where size was the number of 5-year-olds. The second-stage units were schools within sampled PSUs. The third and final stage units were students within schools. The sample consists of children from different racial-ethnic and socioeconomic backgrounds and includes an oversample of Asian, and private school kindergartens. This study estimates reduced-form equations for the health status, for a measure of body mass index (BMI), of U.S. children in the age range of five to twelve. There is a consensus for using ageand gender-specific BMI scores or BMI percentile to study adiposity changes in growing children, particularly with longitudinal studies (Cole et al, 2005; Berkey and Colditz, 2006). Traditional methods for analyzing longitudinal data have been pooled ordinary least square (OLS), random-effects, fixed-effects, or first differencing (Wooldridge, 2002). A recent development of new methods to incorporate the hierarchical structure of data (e.g. nested structure of students within a class or school) includes mixed-effects models (also called hierarchical linear models by Raudenbush and Bryk, 1992) and generalized linear latent and mixed models (GLLAMM) developed by Rabe-Hesketh and Skrondal (2005). In this study, two estimating models are utilized to take full advantage of the hierarchical structure and longitudinal data. An ordered Logit model is estimated with the GLLAMM specification to predict the probability of a child being overweight or obese. A random-effects Tobit model is used to examine factors associated with increased excess weight once a child is already classified as overweight. In the Logit model, the BMI for a given child across the various survey rounds were classified into three categories which are adapted from the Centers for Disease Control and Prevention’s (CDC) weight status categorization for children. The terms “normal”, “overweight”, and “obese” are used to indicate children with a BMI less than the 85th percentile,
3
between the 85th and 95th, and above the 95th percentile in this study. In the Tobit model, the actual measure of BMI is used as the dependent variable, after accounting for the age- and gender-specific cut-off point that distinguish overweight and obese children from normal weight ones. There are several advantages of these new analytic methods. In estimating multilevel models, it is essential to estimate the variance component because there is an assumption that the errors within each randomly sampled cluster are likely to be correlated. With the hierarchical nature of the dataset used in this study, it is important to incorporate random effects into the specification of models estimated (Hox and Kreft, 1994). The fixed-effects term refers to the overall expected effect of households’ and students’ demographics and parent’s time on children’s BMI and obesity status (normal, overweight, or obese). The random-effects give information on whether or not these effects differ between children (level-2) and between schools (level-3)3. In addition, GLLAMM is a class of multilevel latent variable models where a latent variable (common factors or random effects) can be assumed to be discrete or to have a multivariate normal distribution. It is generalized in the sense that it can incorporate different types of response model (e.g. continuous, ordinal, rankings, or mixed responses), with latent variables that can vary at different levels of a hierarchical/multilevel dataset.
2. Analytical framework First an application of the household production model (HPM) theory is summarized. Following is a sketch of the derivation of reduced-form equations for weight status from the theoretical model. The two estimating models applied to the longitudinal data are then discussed.
3
Level 1 is defined as within students across data collection rounds (over time).
4
2.1 Theoretical model The standard economic framework for analyzing children’s outcomes such as health, nutrition intakes, and educational attainment has utilized the HPM model, which incorporates the time dimension into the full budget constraint that households face when maximizing utility. The economic theories of household allocation and children’s outcomes can be categorized into two major strands, unitary and collective/bargaining model4. In the context of this study, we assume a unitary model is sufficient to model the effect of parental work time on childhood obesity for three reasons. First, the traditional conceptualization of separate roles for each parent in a household has diminished over the last several decades in the U.S. Although the gap between the time employed men and women spend in household tasks and childcare still exists, it has decreased dramatically and experts predict convergence (Bond et al, 1998). Therefore, a households’ decision on individual time and resource allocation between parents is viewed not as important as between parents and children within a household. Secondly, unlike developing countries where research has shown mother’s education and work for pay have a substantial impact on children’s health, U.S. parents (male and female) are assumed to share a common view of what is a “healthy weight” and what is not. Thirdly, the age group of interest is 511, which could be viewed as “passive children” since they still mostly depend on their parents for decisions about food consumption and physical activities, both of which have a direct effect on BMI5. A household maximizes utility U[H , X , tl ;ϕ ] which is a function of family members’ health H, consumption of market goods and services X, and leisure tl, subject to a budget constraint
4 5
See Schultz (1999) for a more detailed review of the literature on the development of household models. See Bherman et al (2005) for a thorough discussion and definition
5
pX = wt w + ν where p is exogenous prices for the market basket of goods X; a time constraint T = t l + t w = t Z + t H + t w where time is divided between leisure and market work; leisure is the sum of time used to produce health outcome t H and other household-produced goods t Z . The household production functions are Z = f [ X , t z ] and H i = h[Z , t H ; φ ] where φ is a child’s characteristics.
In this framework, parents’ time and income are pooled together in the full budget constraint in maximizing their household utility, which is also characterized by household preferences ϕ . Specifically, total available time is optimally allocated across outside work, home production, and leisure, which in this case is assumed to not be separable from the home production function6. The
solutions derived from this utility maximizing problem are demand functions for the market goods X * [ p, w, I ; ϕ ] and time used to produce healthy children t H* [ w, p, T , ϕ ,φ ] . By substituting these into the health production function, we can derive an equation for health production
H i * = f [ Z * , th*;ϕ ,φ i ] . Therefore, household total time and income is important in estimating the health outcomes for children in a unitary household model. This framework has been adapted for estimating a wide range of children’s outcomes (Rosenzweig and Schultz, 1983; Senauer and Gracia, 1991; Variyam et al, 1999).
2.2 Reduced form estimating functions The reduced-form health demand function resulting from the above theoretical framework has the general form H = f [ w, p , I | X , ε ] where H is the weight status (either BMI scores or weight status categories), p is a vector of exogenous prices, w is wage rate, I is the household full income, X is a
6
That is, we assume without loss of generality that there is a substitution effect between housework, leisure, and childcare in parental time at home. In other words, we focus on the relationship between time used to work for pay and time used for home production, rather than elements of time used for home production. Besides, it is subjective to identify certain elements of childcare, particularly secondary childcare activities such as playing with a child as being “work” or “pleasure”.
6
vector of observable household and child’s characteristics, and ε is a random variable representing unobserved household and child effects. Theoretically, we can also utilize the optimal results from the unitary household production framework to evaluate the health production function H * = f [ Z * , t Z* ; ϕ , φ ] where H* is the weight status; Z* - a vector of home-produced goods – a solution of the utility maximizing problem described above; t Z* is the optimal time used to produce Z; ϕ is a vector of household factors, and φ is child characteristics. Both of these functions are evaluated in this study.
2.3 Econometric estimating models There are two estimating methods in this study; mixed-effects ordered Logit and random-effects Tobit. We chose order Logit (with three categories of BMI) over the mixed model (with continuous BMI as dependent variable) for two reasons. First, although the underlying BMI is continuous, what matters most is whether children are below or above a certain level of BMI because it indicates whether they are healthy or not. Secondly, in terms of prevention policies, excess weight gain is a more important factor compared to incremental weight gains, which might coincide with growth rates, even after we adjust for age and gender. Therefore, a primary obesity-prevention approach emphasizes efforts that can help normal-weight children maintain that weight and help overweight children prevent further weight gain (Koplan et al, 2005). In addition, a Tobit model is used precisely for the latter reason, to evaluate factors that contribute to the excess weight gain in overweight children.
2.3.1 Mixed-Effect Ordered Logit using GLLAMM The estimating equation is as follows Logit{Pr(y itk = (3categories) | xitk , ς ik( 2 ) , ς k( 3) )} = β 0 + βx
itk
+ ς ik( 2 ) + ς k( 3) + ε itk
0 if y *itk =" normal" where yitk = 1 if y*itk = " overweight" 2 if y* = " obese" itk for ith child at tth round in kth school 7
ε itk has a logistic distribution with variance π 2 / 3
ς ik( 2 ) | x
itk
, ς k( 3) ~ N (0,ψ ( 2 ) ) : a random intercept varying over students (level 2)
ς k( 3) | x
itk
~ N (0,ψ ( 3) ) : a random intercept varying over school (level 3)
Assume that ς ik( 2 ) and ς k( 3) are independent from each other and across schools, and ς ik( 2 ) is independent across students.
2.3.2 Tobit The stochastic model underlying Tobit may be expressed as follows: n BMI i = Max L, β 0 + ∑ β i xi + ε i i =1
Where BMIi is the continuous BMI measure for child ith; L is the limit observation on BMIi being the cut-off point between normal weight and overweight children; and Xi is a set of vectors that include socio-economic, environmental, and endowment factors. Calculations in the Tobit model assume a Tobit index I where I = X ' β . If I falls below a threshold level I* (85th percentile of the weight for age growth chart), the weight status of a child is considered to be zero, that is, the child is not overweight. Therefore, the expected value of BMIi, E(BMIi), is defined as:
E(BMIi) = 0 if I < I* E(BMIi) = I – I* ≥ 0 if I ≥ I* The expected value of BMIi is the average value of BMI for a child, weighted by the probability that the ith child is overweight. The effect of a change in any independent variable on E(BMIi) can be determined as
∂E (BMI i ) = P × βˆ i where P is the probability that BMIi > 85th , and βˆ i is the non∂x i
normalized estimated coefficient.
{
}
x' β The probability of a child being classified as overweight is Prob Yi > 85th | β i xi = F σ
. 8
3. Data and Variables This section describes the set of variables used in the empirical analysis. Descriptive statistics are given in table 1, with a description of each variable. Table 1 reports the means of the quantities and qualities of parental time, environmental, and endowment factors for each category of child’s BMI. Columns (1) and (2) list names and description of variables. Columns (3) to (6) display the mean and standard deviations of all variables in the sample used in this study, overall and by categories of normal, overweight, and obese. The last two columns (7) and (8) describe the significance of pair-wise t-tests between normal and overweight, and between normal and obese children, respectively. These descriptive statistics are shown for the sample used in all regressions in this study, which has 24,512 observations of children over the four rounds of data. For sample attritions which include nonresponse and change in eligibility status over time, see Tourangeau et al (2006) for more details on data collection procedures and comparative statistics.
3.1 Dependent variables The composite variable BMI from ECLS-K5 longitudinal dataset is used as a dependent variable for both models. Although the height and weight components of this composite variable were measured directly by assessors, data entry and coding mistakes might still occur, resulting in some unreasonable numbers such as under 10 or over 40 for children in the age group between five and twelve. Therefore, 30 observations with BMI less than 10 or over 40 were omitted from all four rounds of data. This study uses the terms “normal”, “overweight”, and “obese” to indicate children with BMI less than 85th percentile, between 85th and 95th, and above 95th percentile, respectively7. Therefore, there are three
7
According to the Centers for Disease Control and Prevention’s weight status categorization, children with BMI that rank less than 5th percentile of the growth chart is considered “underweight”; between 5th and 85th percentile BMI are “healthy weight”; between 85th and 95th percentile are “at risk of overweight”; and above 95th percentile are overweight.
9
categories of weights that are used in the ordered Logit model. The cut off points are age- and genderspecific. In the Tobit model, BMI is left-censored for the “normal” weight category (e.g. while reported BMI measures are used for “overweight” and “obese” children.
Table 1: Variables and Descriptive Statistics with Pair-wise t-test (1)
(2)
Variable
Definition
ConBMI
Continous BMI
P1HRS P2HRS CCHOME CCOHOME CCCENTER CCOTHER WOKEARLY WARMCL EVENG PCHINVOL BothPA SingleM SingleD OtherPA NUMSIB PARED IN15C IN30C IN50C IN75C
Person 1 total working hours Person 2 total working hours Childcare at child's home Childcare at someone else's home Childcare at centers Other arrangments of childcare Mother work between childbirth and Frequency of warm close time together Number of evening meals together Index of parent-child involvement Live with both biological parents Live with biological mother only Live with biological father only Live with adopted parent(s) or guardian(s) Number of siblings child has Parents' education Household income,
